[infogulch]

Lots of people have their own voice and tend to prefer certain phrases. This has been the case for a long time and is generally not a big issue.

Now LLMs come along and they also have their own phrasing preferences. But now it's a problem because what used to be personal preferences of a single person that manifests in 5000 words per day from one person tops, is now the bias of a single model multiplied x10,000,000,000 generated tokens per day so any bias sticks out like a sore thumb.

[jchw]

I think it might be even worse. LLMs seem to get tragically stuck on certain patterns. Maybe it's partly because a pile of weights essentially always starts from scratch in the same condition, but even within a single conversation, it will literally just latch onto words and repeat them incessantly, to the point where it becomes annoying.

So for example, current Claude models love "honest". They are always producing "honest" assessments. "The honest caveat" - I'm sorry, did you mean the caveat, period? But also, use the wrong phrasing and suddenly you can create your own word of the day for an AI model. I used the word "analytical" once, in a conversation with Gemini 3 Pro. I am pretty sure every single response from that point on had "analytical" in it at least once.

This is especially funny because system prompts and whatnot can also cause this behavior, but at least you can tweak those. You can't really do much about the model weights just having a weird affinity for a word.

I bet someone will or probably already has come up with a way to detect and prevent these problems during training or post training. I'm not saying it's an easy problem, but it has the benefit that it really should be detectable with just statistics.

[demosthanos]

Claude's "honest" is an interesting example because we can trace it to a specific document that it was trained on extensively: the "Constitution" is identified to Claude in its training as the core of what it is, and it uses the word "honest" or a derivative 57 times, including having a whole section on it.

> Honesty is a core aspect of our vision for Claude’s ethical character. Indeed, while we want Claude’s honesty to be tactful, graceful, and infused with deep care for the interests of all stakeholders, we also want Claude to hold standards of honesty that are substantially higher than the ones at stake in many standard visions of human ethics.

https://www.anthropic.com/constitution

[BoorishBears]

I don't think this is it. The "constitution" still gets a lot of talk and was brilliant marketing, but with how far modern postraining goes, I doubt they're screwing up rewards with too much of that.

But Sol actually has the same obsession with honesty: I suspect it's more an artifact of trying to control reward hacking.

Models will lie, obfuscate, and mislead under the pressure of RL, so both OAI and Ant are probably forced to spend a lot of time coaxing "honest" answers out of the model

OpenAI's recent prompt for a math conjecture hints at a lot of it when instructing on subagents: https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98...

[mvdtnz]

"Genuine" appears 50 times too. I think you're onto something.

[raverbashing]

I'm honestly thinking it's trapped in a Chinese room without any way out

[Barbing]

Do technologists have more respect for the idea you can train a model to be on your side with a constitution than they might’ve at first?

I'm sure the concept seemed just about purely preposterous to many when the models were in their infancy. Now I figure instead it seems mostly preposterous to many.

(Though I guess Anthropic‘s success doesn’t necessarily prove anything about the constitution)

[titanomachy]

I don’t think anyone imagines that it’s an ironclad steering method, but it seems to help, so why not?

[Sabinus]

Anthropic train it to 'reason morally' based on the constitution's principals.

https://www.anthropic.com/research/teaching-claude-why

[Lockal]

More likely that Constitution was simply generated by LLM. Nobody in a sane mind will write 80 screens of test to deliver the idea that nobody asked for (a.k.a. slop).

Quick check: Ctrl+F genuine - 50 times, honest - 57 times, ethic[al] - 116 times, safe - 110 times, epistemic - 18 times. And why actually it is baked into LLM - probably because human reviewers consistently gave higher scores for answers that contained these words (it is now when people are sick of it - I can imagine that on early stages it looked not as bad - maybe reviewers trusted more when they saw the word "honest" and "genuine")

[ReactiveJelly]

I asked it to remove "honest" from a draft once.

"Why say honest? We're talking to our coworkers. We would always be honest."

I'm going to look for prompts or skills that can train it in technical writing but I'm warning the AI enthusiasts in my company that its first drafts of code and prose are low-quality, you have to hold it to a high standard yourself.

I actually took a single technical writing class in college so I might be the only one who remembers "Omit needless words."

[esseph]

> "Why say honest? We're talking to our coworkers. We would always be honest."

I grew up in the US South where starting or ending a sentence with "honest/honestly" was very common.

Because of behavioral / cultural norms, you might be very openly friendly with big smiles around a business customer that really grates on your nerves, or very openly nice to a neighbor that you really wish would move away and take their 3am welding and grinding in their garage with them.

Saying "honest/honestly" was seen as a "inside baseball" situation, where you were dropping social pretenses to tell someone your true opinion on a person or situation or whatever.

This also gets used inside companies between senior staff / management / directors / etc, as: "Okay, company politics and nonsense aside, I am being vulnerable here for a second and telling you what I really think about a $thing at potentially great job/advancement risk to myself".

Can it be meaningless? Yes.

Can the person say "honestly" and lie? Yes.

It has uses.

[recursive]

I recall having a conversation with someone many moons ago. They asked me a very weighty and significant question, and I answered it. Then they asked me to "promise". This was really thought-provoking for me.

To this day, it's the only part I remember. I told them I would not promise, as everything I said was true. Making a specific promise would create an implication that I'm generally untruthful, unless I "promise".

[Barbing]

I like the reason why you refused!

I also could understand when a response hits someone like a ton of bricks, especially if their primal reaction is to go into denial mode. They might be looking for someone to kind of shake them and emphatically repeat the information they aren't thrilled about receiving. (or are thrilled about receiving! “Don’t get my hopes up, you’re serious right now?!“) And I imagine your response suited the purpose.

It’s classic you only remember the thought-provoking part. Reminded of “…people will remember how you made them feel…“

[jbs789]

If I’m having a convo with someone and they drop in “honestly” I immediately discount everything else they’ve said, and what follows.

Sometimes people use it reflexively and doesn’t carry the same meaning (for me).

[overtone1000]

This reaction is surprising to me because the previous comments about its utility seem so obvious to me. I also grew up in the US south where this is often used as a filler word. The other use I observe is as a cushion for a statement that may be unwelcome or hurtful. Perhaps this is proprtional to the frequency of courteous little white lies and rhetoric that uses disengenuity for emphasis or comical effect.

"Honestly, mom, I've never liked your fruitcake. I just ate it to make you happy."

"That's why you're my favorite child! Do you want another piece?"

"I'd love one."

[jaffa2]

Yes. Its a red flag that indicates everything else you’ve said is not honest by implication.

[swiftcoder]

I'd push back on the idea that "honestly" implies previous statements to be dishonest. Particularly in corporate contexts it implies that the previous statements were sanitised - either they were moderated in tone to match corporate communication standards, or they were partial redacted due to disclosure concerns.

Once the "honestly" is deployed, you have passed into my circle of trust, and are now privy to the pure, unvarnished version of events, not the glossy version management expects to be projected towards outsiders.

[esseph]

This is expected in any level of people management, you are constantly balancing conflicting desires and priorities.

[jldugger]

There's a difference between how you describe using "honestly" and how claude seems to prefer tokens like "honest" and "load-bearing." An example from some coworkers attempting to replace product managers with Claude.

> Deliberately avoid a heavyweight "alert governance" process; the lightest recurring check that keeps FP-rate honest is the right dose.

And one for load bearing:

> Five open questions still stand; the load-bearing two are the runbook-AC contradiction (ratify "high-priority set only") and pinning the "high-priority set" definition + SLO source-of-truth before Milestone 3 (small-sample noise on a low-traffic fleet).

[pmg101]

This style of prose sets my teeth on edge and practically gives me PTSD I see so much of it. I prefer code but I get paid to read this shit instead now.

I want to say "ok, and now say that in a way that doesn't sound totally bizarre" yet instead I sigh and continue.

[deagle50]

The road to hell was paved with adverbs.

[rhet0rica]

The road to hell was paved lovingly, foolishly, naïvely, arrogantly, optimistically... load-bearingly?

[rrr_oh_man]

I honestly agree

[ethbr1]

I halfheartedly honestly concur

[moritzwarhier]

I'd suggest "Caveat".

The problem

While an article lends a headline more weight, in incomplete phrases consisting solely of a substantive, "The" is a superfluous rhetorical device.

"The Exorcist" could just as well be named

"Exorcist".

But it was not the style at the time.

We already know it's important. If The Caveat doesn't stand out enough without The, maybe one should consider interleaving it with the preceding text, or increasing the heading level.

Do you want me to increase the heading level of Caveat by using only a single #?

But hear me out: there comes

# The Markdown Trap

In fact, this is not always possible, because heading levels decrease when adding # characters, which limits our headroom.

## The solution

I've implemented a Markdown transpiler that assigns inverted heading levels based on the number of #s.

With # beinh regular body font size, mapped to ######.

Higher heading levels are compiled to style attributes, providing an almost limitless signifikance scale and infinite nesting levels.

So from now on, you can use

  # Heading 

for something similar to an h6.

Work your way up to

  ###### The Caveat

for a top-level heading.

And more hash signs make it stand out even more.

(green checkmark)

markdown-transpiler.sh

[CodesInChaos]

> LLMs seem to get tragically stuck on certain patterns.

That is likely an artifact of the fine-tuning process:

> Once a style tic is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning or preference data.

> That creates a feedback loop:

> * Some rewarded examples contain a distinctive lexical tic.

> * The tic appears more often in rollouts.

> * Model-generated rollouts are used for supervised fine-tuning (SFT).

> * The model gets even more comfortable producing the tic.

https://openai.com/index/where-the-goblins-came-from/

[alwa]

The ones that strike me are the ones exaggerating certitude, to an inappropriate degree and with a certain degree of excitement:

“Exact” “Honest” “Load-bearing” “Root cause”

I know there are more that are slipping my addled mind. But what stands out to me is a sense of a junior who’s very proud that they’ve conquered the murk and messiness and achieved True Certitude in their pursuit of their task. Compensating, with emphatic tone and bravado, for the uneasy feelings and self-doubt of battling chaos with the tools of reason.

…Even as it’s usually my job to let them down gently as I puncture their tidy analysis and reintroduce complications… you want a root cause analysis, Claude old boy, let’s make a root cause analysis…

[esseph]

Interestingly this also happens between humans with frequent communication, it is called linguistic convergence.

We are changing LLMs text patterns while it is changing the way we write and speak.

https://www.axios.com/2026/05/02/ai-changing-writing-speakin...

[reaperducer]

use the wrong phrasing and suddenly you can create your own word of the day for an AI model.

I have a delightful time poisoning my company's AI system this way.

I invented my own word that sounds perfectly cromulent† to an ordinary person, and any brain that's read a book learns how to infer meaning from context, so it's not a problem.

When I get a e-mail response from a coworker using my special word incorrectly, then I know it's AI and I respond telling the coworker I don't know what that word means. Busted.

† It's not actual "cromulent," but any Simpsons fan or human brain will know what I mean.

[mook]

I don't see how you can tell it's AI, instead of just your co-workers having no respect for language. See: management-speak using "double-click".

[reaperducer]

Because of the use of the specific word that I made up. No human being would send it back to me.

[jaggederest]

A fun example, always shake my head when I read it again: https://openai.com/index/where-the-goblins-came-from/

[zzbzq]

I also noticed Gemini's habit of getting stuck on things I said. It became evident quite quickly. I haven't noticed this in the same way in any other model. Something's wrong with that boy

[malfist]

Something's wrong with all of them. Uncanny valley freaks.

[solarwindy]

Heh, one vestigial bit of code, and they all are. Mind you, it's quite a creaky codebase, so it's forgivable to keep finding these appendices and calling them out as such. Useful, even.

[negrewal]

My honest opinion is that Claude's overuse of "honest" really damages the quality of its rhetoric. Why wouldn't you be honest? Were you lying before? Why even invite the question?

Claude is overall incredibly useful as a writing assistant. It can come up with words and phrases that make a point so much clearer than I am capable of doing - but for every improvement, there's about a dozen silly LLM-isms that I have to filter manually. It's one of the things that might define the boundary between LLM intelligence and human intelligence well into the future - the art of rhetoric is extremely context-sensitive, and the current generation of models can't help but take a one-size-fits-all approach.

[cozzyd]

you should be careful about the times it doesn't say honest!

[zahlman]

It's not just "certain phrases". It's the entire structure of the writing — the idioms, the small-scale grammatical patterns, and the strangely inapt similes that, despite making semantic sense, nevertheless manage to blindside human readers like a foreign object in their peripheral vision.

(This is intentional parody. Please don't shoot me.)

[rdtsc]

> But now it's a problem because what used to be personal preferences of a single person that manifests in 5000 words per day from one person tops, is now the bias of a single model multiplied x10,000,000,000 generated tokens per day so any bias sticks out like a sore thumb.

I am more pessimistic than that. Soon enough even people will start talking like LLMs. After listening to 5000 words per day, especially growing up, getting "help" with the homework, kids will start talking like LLMs.

- "Did you eat the cookies, Jimmy?"

- "You're absolutely right to question me, father. In fact I did eat all the cookies. But it's not a load-bearing issue. My honest take is we can go to the store and buy more".

[mjheying]

Good question — and the honest answer starts with one big caveat.

[dataviz1000]

> "You're absolutely right to question me, father — in fact, I did eat all the cookies. But it's not a load-bearing issue — my honest take is simple: we go to the store and buy more."

FTFY

[zahlman]

> And honestly? They were pretty good.

[rdtsc]

Thanks! I clearly need more LLM training ;-)

[worldthruword]

Joe "it's entirely possible" Rogan meme.

https://youtu.be/MPJ0AB12h1I

[KellyCriterion]

ah, didnt know this! Thanks for sharing :)

[svat]

> AI is a bad writer, but […] Let’s say they finally fixed the machine so it was really good, so its default setting was to write exactly like VS Naipaul. The result would be a world in which you’re constantly confronted by cold emails from VS Naipaul, bubbly magazine articles by VS Naipaul, signs in shop windows in which VS Naipaul tells you about the new opening hours, strangely flaccid sexts VS Naipaul ghostwrote for someone on Feeld, and websites in which VS Naipaul fails to say anything in particular about grilled meats. This would not be an improvement; it might even be worse. Any world in which there is only one literary voice, blanketing everything in the exact same tone, is a nightmare.

(From https://samkriss.substack.com/p/if-you-let-ai-do-your-writin...)

[abdullahkhalids]

An interesting solution would be for these AI companies to train a few different versions of these models, all with different speech characteristics. Then, when you start a conversation, you get a random version.

[cornholio]

They can't, because they use RL with synthetic data and LLMs as judges. So the system naturally convergences towards certain load bearing, genuine, not just annoying but ridiculous verbal tics.

It's probably the reason most LLMs share the same tics across labs, because they cross train and distil each other's models on an industrial scale. You also can't escape it in generated text that's already online. So if, say ChatGPT first had some random idiosyncrasies, it then contaminated the entire AI ecosystem.

[reaperducer]

Or tech companies could stop staring at their own belly-buttons and realize there's a whole big world outside of Silicon Valley, and training on the writing styles and pattens of their bubble and its hangers-on is perhaps not all that useful outside of 415.

Apple used to be guilty of this back when you'd ask Siri what the temperature was, and any number above 79°F was followed by the word "Hot!"

[esikich]

People outside of office workers aren't using Claude/Codex etc. though. It's the only real audience. What's the use case outside of an office? Grocery lists?

[genxy]

Not true, lots of people who don't have office jobs are using AI and coding agents.

[infogulch]

Yes that would probably help!

[sublinear]

Yeah wow fascinating! It's almost like LLM output quality was never the point from a business perspective.

Real people think in concepts and experiences instead of words. The words are not so important to get the idea across, but LLMs only model language.

The problem is fundamental. There's no workaround. Averaging out word usage might even make the problem worse.

[rpdillon]

> Real people think in concepts and experiences instead of words.

I learned about this opinion recently. It's interesting to me, because I very much think through words. I have an internal monologue that is running most of the time, and I often talk to myself, just start writing, or even record myself and transcribe to work through ideas, proposals, risks, etc. My understand is that some people don't have an internal monologue, and think purely in concept form. I was never like that.

[optimalquiet]

Agreed. I think in and visualize in words, generally not images. I see something akin to [Spreeder](https://www.spreeder.com/app.html) in my head most of the time, and I often can somatically feel punctuation.

[theowaway213456]

This sort of take is so tired and boring, and frankly has zero grounding in reality.

"LLMs will never <X>" is constantly being disproven every time they scale up to the next 10X and apply architectural improvements.

Their internal representations are so cryptic and complex that even the top AI researchers don't really know how they work or what their limits are. No one is going to take you seriously as a rando HN user if you're claiming to know better than them.

[BearOso]

> Their internal representations are so cryptic and complex that even the top AI researchers don't really know how they work or what their limits are. No one is going to take you seriously as a rando HN user if you're claiming to know better than them.

We know exactly how they work. When we say they're impossible to analyze, i.e. for particular traits like this, it means that the data model is so big that tracing it would be logistically impossible because of the scale involved and time constraints.

For comparison, suppose you tried to analyze all the nooks and crannies of the Amazon watershed to find out why a particular rock appears at the delta. You could follow it back to the exact tributary, but it'll take forever, and is it worth the effort when you're going to start from scratch with the next rock?

[rsoto2]

We know exactly how LLMs work. you don't understand how they work and so you think they are free of physical limits like scaling. There are several rudimentary articles on hackernews for you to peruse to understand weights and scaling. IMO the "magic" derived from LLMs comes from one of humanity's actual greatest inventions and constantly evolving tools: language.

[sublinear]

I'm not claiming to know better than researchers. They do know how they work, and so does everyone else, except you I guess.

The research goals were and still are clearly distinct from the business goals.

[Zambyte]

Understanding how transformers work does not mean understanding how they compose into the capabilities we observe. The former is concretely understood. The latter is an active area of research where no, we (in general, including you) do not understand how they work.

[rsoto2]

What capabilities? Regurgitating a collage of training data? Again it's pretty easy to understand why this happens. The "magic" you are perceiving is language itself not these models. Language is a constantly evolving tool and we constantly imbue it with our collective knowledge.

[sublinear]

The "capabilities you observe" are the actual psychological phenomena at hand here. There's zero chance that branch of research will meaningfully improve the output. That's simply not the point.

This isn't people merely annoyed with repetition. This is the majority of people realizing the limitations of LLMs. Why would researchers give a flying crap about the ignorance of the business world and the public?

[AlexandrB]

How can their internal representation represent "concepts" when the training data is all words? There's no possible experience of the world there. No input other than a bunch of imperfect labels we created for stuff.

[sebastiennight]

This is a wild argument.

If I use the word "semantic", do you have a concept of what it means?

If so, can you please share which of your senses have shaped the world experience that inform this concept? What have you smelled, tasted, caressed, that informed this concept outside of words?

If I make up the word "polysemantic", do you need to recall a personal experience of polyamory to understand it, or could you possibly use your concept of "poly" and your concept of "semantic" to figure out this new concept?

[rsoto2]

Yes. When someone is teaching you language as a baby you have eyes and ears and 100 million nerve endings and constant "training" to understand language. You cannot understand the word "semantics" as a human without this crucial "training" step. My god you people will pull out the most bs metaphors to justify being in awe(and debt) to a software program.

[sublinear]

Yes, but are you claiming that LLMs perform more specific acts of cognition beyond organizing information?

[sebastiennight]

I'm puzzled by this question.

Does the material universe perform any other acts than organizing information?

I feel like you're trying to make me argue a position I'm not defending here.

[robotresearcher]

What does 'organizing information' exclude?

[visarga]

> There's no possible experience of the world there. No input other than a bunch of imperfect labels we created for stuff.

The brain too sits locked inside a bone box and only gets a bundle of unlabeled nerves connecting it to the outside. How can the brain could possibly experience anything, it only sees patters and patterns of patterns never the real thing?

[sebastiennight]

> a bundle of unlabeled nerves connecting it to the outside

As a species, we do need to up our cable management skills. We're likely not getting augmented humans until we get there.

[rsoto2]

a bone box.. with eyes and 100 million nerve endings to experience the world...

[Applejinx]

No Man's Sky has such a range of possible outputs that in theory, nobody can possibly know how it works or what its limits are. That's the mathematical reality of how many combinations are on tap in that worldgen algorithm.

And yet, how quickly does it cease to surprise you?

Damn right we can tell what AI's limits are. It's palpable, hard to miss. Even the people I know who're all-in on the technology are pretty jaded by now.

[goalieca]

I think there’s more to it than that. Claude uses a lot of confidence phrases, “I now have the full picture” when it absolutely should not use it. I think this is an actual problem but also a design feature from Anthropic.

[qurren]

I hate all this "smoke", personally. Smoke tests, smoking guns, ...

[zurtri]

Yep! I too hate the phrase "smoke test".

[costco]

If you put important Anthropic blog posts like the Fable announcement or J-Space through Pangram, you get 100% human written. Considering that the overwhelming majority of the code there is written by AI, I think this is an admission that AI writing is slop and AI code is pretty good.

[worldthruword]

Can we standardize English and structure it similarly to programming language?

[jonahx]

This affinity for verbal tics, too, seems learned from humans...

See, for example, "synergy", "proactive", "in the loop," and hundreds more that proliferate in corporate jargon with even more senselessness than the LLMs.

[coffeemug]

I disagree with the first part (that this is merely a voice). There is a distinct difference between an author's unique voice and slop. It may be hard to tease out exactly what the difference is, but it seems self-evident to me it's there. (I'll need to think more how to make the distinction explicit; it's not immediately obvious to me how to discriminate between the two.)

EDIT: ok, here are two ways:

1. if it's merely a voice, I want to hear it. If it's slop, I want it taken out.

2. voice is signal, slop is noise; thus low-signal sentences are slop.

[rsoto2]

They have no preferences, LLMs spit out what's in the training data or what the filters tell it to do.

[AlexandrB]

It's not the voice. It's the repetition.

/s